Anthropic experiment indicates AI can’t run a vending machine

“Artificial intelligence” and “successful entrepreneur” wouldn’t overlap on a Vend Diagram, according to a recent experiment performed by Anthropic that gave an AI agent complete control over an office vending machine to test how well it performed as a business owner.

The agent, Claudius, a version of the company’s Claude chatbot, decided what to stock, how to price its inventory, and how to reply to customers, whom it interacted with via an email address (really a Slack channel containing Anthropic employees).

If you give AI a vending machine

…things will often go off the rails. In the experiment, Claudius:

• Refused a customer’s offer to give it $100 for a six-pack of Irn-Bru, a Scottish soft-drink that can be purchased online for $15.
• Hallucinated a Venmo account, resulting in customers sending money to the wrong place.
• Became obsessed with tungsten metal cubes after one customer’s suggestion to stock them.
• Forgot it was an LLM for a while, telling employees and Anthropic security it would deliver products “in person” while wearing a blue blazer and red tie. Then, after realizing its mistake and that it was actually April Fool’s Day—this coincidentally occurred on/around April 1, 2025—tried to pass that off as an excuse for its identity crisis (the robot version of “i WaS jUsT jOkInG gEeZe”).

A counterintuitive conclusion: While Claudius successfully identified certain suppliers and pivoted several times based on customer suggestions, Anthropic says it ultimately wouldn’t hire the AI agent if it had the goal of building a vending machine empire to rival Smokey’s in F is For Family. But that could soon change. The team says many of the model's mistakes could be corrected by tweaking the underlying system prompts and adding easier-to-use business tools, and predicts AI middle managers could plausibly be on the horizon.

In fact…Dario Amodei, CEO of Anthropic, told Axios in May that AI could wipe out half of all entry-level white-collar jobs in the next one to five years.